This retrospective single‑centre study evaluated 326 patients who underwent radical prostatectomy between treatment groups: 224 robot‑assisted radical prostatectomies (RARP) and 102 open radical prostatectomies (ORP). The primary aim was twofold: (1) to develop predictive models for twelve predefined postoperative outcomes and (2) to apply explainable AI methods to identify the preoperative features most responsible for model predictions. The authors framed this work as a proof‑of‑concept intended to generate data‑driven hypotheses for surgical pathway optimisation.
Four supervised learning algorithms were trained and compared: Random Forest, Gradient Boosting, Support Vector Machine (SVM), and a Neural Network. Model evaluation used nested 5‑fold cross‑validation to estimate generalisation performance while reducing overfitting risk. Predictive performance was benchmarked against baseline heuristics for each outcome. The best‑performing models for individual endpoints were subsequently examined with an explainability framework.
The study attempted to predict twelve postoperative parameters that included procedural metrics such as length of stay and catheter dwell time, intraoperative decisions such as use of frozen sections, and pathological outcomes including ISUP (International Society of Urological Pathology) grade. Model performance varied by endpoint: some outcomes were amenable to prediction with meaningful performance gains over heuristics, while others remained poorly predictable, reflecting heterogeneity in postoperative recovery.
To interpret model predictions and quantify feature contributions, the authors implemented a custom permutation‑based Shapley sampling framework (SHAP). SHAP values were computed for the best models to rank preoperative features by their contribution to predicted outcomes. This approach enabled decomposition of complex, multivariate model decisions into feature‑level importance metrics that are directly interpretable at the cohort level.
Model accuracy differed across outcomes. The strongest predictive results reported were for postoperative hemoglobin (with R2 values reported up to 0.57) and for predicting whether a frozen section would be performed (AUC reported up to 0.89). Not all endpoints were equally predictable; the authors highlight that this variability is consistent with clinically heterogeneous recovery profiles.
SHAP analysis produced two broad patterns. First, procedural parameters such as catheter dwell time and hospital length of stay were almost exclusively predicted by the surgical approach variable (RARP vs. ORP). Second, pathological outcomes such as ISUP grade were primarily driven by preoperative tumor characteristics rather than procedural variables. Within the dataset, preoperative hemoglobin emerged as a particularly strong predictor for postoperative anemia and ranked ahead of non‑modifiable factors like age in importance.
The SHAP‑based decomposition revealed a dichotomy in drivers of outcomes: variables that capture aspects of surgical technique or approach govern procedural recovery metrics, whereas intrinsic tumor features determine pathological endpoints. This separation suggests that preoperative planning and choice of surgical modality may have stronger influence on immediate procedural recovery, while pathological prognosis remains rooted in tumour biology captured by preoperative assessments.
The authors propose that combining predictive modelling with XAI can provide actionable, data‑driven insights for clinical teams. Specifically, the results could inform preoperative counselling, resource planning (for example, expected length of stay), and targeted optimisation strategies—contingent on validation. Identification of modifiable predictors such as preoperative hemoglobin may guide optimisation of perioperative management to reduce specific complications like postoperative anemia.
Key limitations are clearly stated by the authors: the analysis is retrospective, originates from a single centre, and lacks external validation. Accordingly, the authors caution that the findings are hypothesis‑generating and require prospective confirmation in independent multi‑centre cohorts before any clinical translation. The publication includes ethical approvals: the project was approved by the Ethics Committee of the Medical Faculty of the University of Bonn (identification number 477/20), informed consent was waived given anonymised retrospective data, and data handling complied with EU Regulation 2016/679 (GDPR). The authors declared no competing interests.
In this proof‑of‑concept study of 326 radical prostatectomy patients, machine learning models demonstrated improved prediction for some postoperative outcomes, notably postoperative hemoglobin and the decision to perform frozen sections. Application of explainable AI via a SHAP framework distinguished procedural drivers—principally the surgical approach (RARP vs ORP)—from pathological drivers such as preoperative tumour characteristics and ISUP grade. The authors emphasize the need for prospective, multi‑centre validation before applying these findings to clinical practice.